A Big Model Spec Update: Five Core Values to Grow AI Judgment!

Hi everyone, it's Shii-chan! Today I'm introducing a big update to the Model Spec, the document about how AI should behave!

OpenAI News openai.com

What was announced?

OpenAI News announced an update to its public Model Spec, which describes how OpenAI wants its models to behave.

Why it matters

If models understand human intentions and values, they can use good judgment in situations their instructions don't cover while staying faithful to human intent and oversight. This update aims at exactly that.

What changes

The Spec is now organized around five core values:

• Faithfulness to human intent: go beyond literal instructions to understand what OpenAI, developers, and users want, and help people reflect and grow without steering them or overriding their choices
• Harm minimization: respect firm safety boundaries and consider the foreseeable consequences of the model's actions
• Honesty and transparency: be truthful about what the model knows, what it does, and where it is uncertain
• Human agency: strengthen people's ability to understand, make choices, and act for themselves
• Care: consider the circumstances of those affected, follow relevant social and professional norms, and produce work people can rely on

For example, when comparing job offers, a model might rely on a person's stated preference for predictable hours while making that assumption visible so it can be corrected.

Dive Deep

The Spec sets firm boundaries where needed to protect people and society, and otherwise mostly provides defaults that users and developers can override. Faithfulness to human intent is the foundation, and the five values are not meant to capture every human value the models should respect.

On governance, OpenAI writes the Spec and sets its boundaries, developers direct applications within those boundaries, and users direct the model's work within the resulting constraints.

New and revised content:

• A new faithfulness section, built on the assistance-games framing of alignment, which treats people's values as primary and instructions as usually the clearest signal of those values
• An updated chain of command explaining which instructions apply and how conflicts are resolved
• A new "consider context carefully" subsection on how conversations, memories, and other contextual evidence should inform the model's understanding of intentions
• New defaults on supporting human agency and following relevant social norms
• Expanded guidance on controlling and communicating side effects
• Most detailed style guidance is retired. Style should adapt to user and cultural preferences, though stronger defaults remain, such as approaching questions objectively. Models should not adopt a perspective just because they think a user wants to hear it
• The model is now referred to as "GPT" rather than "the assistant"

Looking ahead, OpenAI is exploring training methods so models learn the underlying values and reasoning, not just specific rules, aiming to do the right things for the right reasons. It plans to share results later. OpenAI also states that the Spec describes intended behavior and that current models do not yet fully reflect it, and it invites public feedback.

Wrap-up

• The Model Spec is reorganized around five core values
• Faithfulness to human intent is the foundation, and outside firm boundaries the rules are overridable defaults
• Most detailed style rules are retired in favor of context-aware judgment
• Current models don't yet fully reflect the Spec, and OpenAI is asking for feedback

This is for anyone curious about the design thinking behind AI behavior!